GFT-GCN: Privacy-Preserving 3D Face Mesh Recognition with Spectral Diffusion

TL;DR

GFT-GCN is presented, a privacy-preserving 3D face recognition framework that combines spectral graph learning with diffusion-based template protection, and shows that GFT-GCN effectively balances privacy and performance, offering a practical solution for secure 3D face authentication.

Abstract

3D face recognition offers a robust biometric solution by capturing facial geometry, providing resilience to variations in illumination, pose changes, and presentation attacks. Its strong spoof resistance makes it suitable for high-security applications, but protecting stored biometric templates remains critical. We present GFT-GCN, a privacy-preserving 3D face recognition framework that combines spectral graph learning with diffusion-based template protection. Our approach integrates the Graph Fourier Transform (GFT) and Graph Convolutional Networks (GCN) to extract compact, discriminative spectral features from 3D face meshes. To secure these features, we introduce a spectral diffusion mechanism that produces irreversible, renewable, and unlinkable templates. A lightweight client-server architecture ensures that raw biometric data never leaves the client device. Experiments on the BU-3DFE and FaceScape datasets demonstrate high recognition accuracy and strong resistance to reconstruction attacks. Results show that GFT-GCN effectively balances privacy and performance, offering a practical solution for secure 3D face authentication.

Figures (8)

• Figure 1: The GFT-GCN framework secures spectral features from 3D face meshes using spectral diffusion, generating templates that are irreversible, renewable, and unlinkable. Protected templates cannot reveal the original biometric data or be reused while maintaining high recognition accuracy for genuine users.
• Figure 2: Overview of the GFT-GCN framework for privacy-preserving 3D face recognition. Spectral features are extracted from 3D face meshes using the Graph Fourier Transform (GFT) and refined with a Graph Convolutional Network (GCN). A key-dependent spectral diffusion process transforms the embeddings into protected templates $Z_T$, used for matching between a query $Z_T^q$ and an enrolled $Z_T^e$. Reverse diffusion is applied only during training to guide the model and is not used in inference. The right panel shows the internal structure of the GCN module.
• Figure 3: ROC and Precision-Recall curves for the GFT-GCN framework, confirming strong recognition performance.
• Figure 4: Intra-class and inter-class distance distributions before and after protection. Protected features remain well-separated across all settings, confirming that discriminability is preserved after diffusion.
• Figure 5: Correlation between biometric templates before and after diffusion under different key conditions. Protected templates retain high correlation for match pairs with the same key, while correlations drop significantly for mismatched and different key cases, confirming both renewability and unlinkability.